# Lead AI-native delivery **Agentic Engineering Field Curriculum · Edition 2026.08 · Reviewed 2026-08-10** Redesign the operating system around higher generation speed while protecting strategic judgment, independent assurance, and outcome ownership. Level: practitioner · Designed for Executive, Engineer, Risk, Operations · Estimated study: 9 hours · 6 finite stages Canonical page: https://aisdlc.ai/agentic-engineering/learn/lead-ai-native-delivery ## 01 · Measure the system, not the activity **Driving question:** Which numbers show that delivery improved, and which only show that people and agents generated more? Build a measurement set that separates local generation activity from delivered outcomes, stability, learning, and human cognitive load. ### Learning objectives - Distinguish activity counts from throughput, stability, and outcome signals. - Pair every speed measure with a stability, quality, or human-load measure. - State the failure each measure is meant to detect and the blind spot it carries. - Reject measures that no named person will act on. ### Study sequence 1. **[Velocity sickness](/agentic-engineering/velocity-sickness)** — read, 14 min. Name the failure this path exists to prevent before choosing a single measure. 2. **[Compare activity measures with system measures](/insights/own-the-speed)** — compare, 20 min. Place each candidate measure against what it can detect and what it will miss. 3. **[Assemble a balanced measurement set](/adoption-workbench)** — practice, 35 min. Turn candidate measures into one defined, decision-linked set. ### Field exercise Build a measurement set for one delivery group that reports delivered outcomes, stability, learning, and human load alongside generation volume. **Deliverable:** A measurement set with definitions, sources, paired counter-measures, and a blind-spot register. **Independent challenge:** A reviewer outside the delivery group checks whether any measure in the set could rise while delivered outcomes, stability, or human load got worse. **Evidence to retain** - Measure definitions - Collection sources and dates - Paired counter-measures - Blind-spot register - Decision map **Connected knowledge** - [Velocity sickness](https://aisdlc.ai/agentic-engineering/velocity-sickness) - [Organizational absorption capacity](https://aisdlc.ai/agentic-engineering/organizational-absorption-capacity) - [Agentic software factory](https://aisdlc.ai/agentic-engineering/software-factory) - [Observability](https://aisdlc.ai/agentic-engineering/observability) - [Evidence engineering](https://aisdlc.ai/agentic-engineering/evidence-engineering) **Connected Insights** - [Own the speed](https://aisdlc.ai/insights/own-the-speed) ## 02 · Find the absorption constraint **Driving question:** Where does generated work stop moving, and what does that queue cost? Map the decision, review, verification, integration, release, and operating queues that determine how much generated work becomes delivered value. ### Learning objectives - Trace one work item from request to operated outcome and record every wait state. - Record wait time separately from active time at each handoff. - Count rework loops as separate passes through a queue. - Name the single constraint that limits verified throughput before proposing capacity anywhere else. ### Study sequence 1. **[Organizational absorption capacity](/agentic-engineering/organizational-absorption-capacity)** — read, 15 min. Define the capacity model this module measures a real value stream against. 2. **[Compare a production system with a queue amplifier](/agentic-engineering/software-factory)** — compare, 20 min. Separate repeatable production from generation that only lengthens downstream queues. 3. **[Trace work items end to end](/adoption-workbench)** — practice, 40 min. Produce the timeline evidence a constraint claim must rest on. ### Field exercise Trace three recently completed work items through every queue between request and operated outcome, then locate the constraint that limits verified throughput. **Deliverable:** A value-stream map with wait time, rework counts, and one named constraint supported by the traced timelines. **Independent challenge:** The owner of each mapped queue checks the recorded wait times and rework counts against their own records and challenges any figure they cannot reproduce. **Evidence to retain** - Work-item timelines - Queue inventory - Wait and active time split - Rework log - Constraint statement with its basis **Connected knowledge** - [Organizational absorption capacity](https://aisdlc.ai/agentic-engineering/organizational-absorption-capacity) - [Agentic software factory](https://aisdlc.ai/agentic-engineering/software-factory) - [Human accountability](https://aisdlc.ai/agentic-engineering/human-accountability) - [Durable execution](https://aisdlc.ai/agentic-engineering/durable-execution) - [Observability](https://aisdlc.ai/agentic-engineering/observability) **Connected Insights** - [Own the speed](https://aisdlc.ai/insights/own-the-speed) ## 03 · Redesign work and attention **Driving question:** How should work be sized, sequenced, and packaged so human judgment lands where it changes the outcome? Set batch size, work limits, decision packets, and protected review time from measured capacity so human authority stays real under higher generation speed. ### Learning objectives - Size work so one unit can be reviewed and verified inside a single attention window. - Set work-in-process limits from measured review and verification capacity. - Package a decision so the accountable person can rule on it without reconstructing the work. - Schedule protected decision and review time as a standing commitment. ### Study sequence 1. **[Human accountability under delegation](/agentic-engineering/human-accountability)** — read, 14 min. Establish what a human decision must contain before work is redesigned around it. 2. **[Write one decision packet](/enterprise-aisdlc)** — practice, 30 min. Turn a live decision into a packet a named person can rule on without a walkthrough. 3. **[Compare batch sizes against review capacity](/agentic-engineering/organizational-absorption-capacity)** — compare, 20 min. Derive a defensible work limit from measured capacity rather than preference. ### Field exercise Redesign one group's work intake so batch size, work limits, and decision packets match the review and decision capacity measured in the previous module. **Deliverable:** A revised intake design with batch rules, a work-in-process limit and its basis, a decision packet template, and a protected review schedule. **Independent challenge:** A reviewer who did not write the packet rules on one sample packet using only its contents, and reports any question they could not answer from it. **Evidence to retain** - Batch sizing rule - Work-in-process limit and its basis - Decision packet template - Completed sample packet - Protected time schedule **Connected knowledge** - [Human accountability](https://aisdlc.ai/agentic-engineering/human-accountability) - [Organizational absorption capacity](https://aisdlc.ai/agentic-engineering/organizational-absorption-capacity) - [Intent engineering](https://aisdlc.ai/agentic-engineering/intent-engineering) - [Goal and exit condition](https://aisdlc.ai/agentic-engineering/goal-exit-condition) - [Evidence engineering](https://aisdlc.ai/agentic-engineering/evidence-engineering) **Connected Insights** - [Own the speed](https://aisdlc.ai/insights/own-the-speed) - [Evidence before autonomy](https://aisdlc.ai/insights/evidence-before-autonomy) ## 04 · Scale verification and evidence with generation **Driving question:** What has to grow alongside generation capacity so more output does not quietly mean less assurance? Plan independent verification and release evidence that grows with generation capacity, with named ownership and a stop condition when capacity is exceeded. ### Learning objectives - Separate builder self-checks from independently owned challenge. - Size verifier ownership and evidence capture against expected release volume. - Define the evidence each release must carry before it can advance. - Set a stop condition tied to a measured verification backlog. ### Study sequence 1. **[Verification is a plane, not a phase](/insights/verification-is-a-plane-not-a-phase)** — read, 15 min. Establish where challenge belongs before sizing how much of it a delivery group needs. 2. **[Compare self-review with independent challenge](/agentic-engineering/multi-agent-verification)** — compare, 20 min. Locate the independence boundary a capacity plan has to respect. 3. **[Size the verifier plane against release volume](/assurance-case)** — practice, 35 min. Convert the independence boundary into named ownership and a countable capacity assumption. ### Field exercise Plan the verification and evidence capacity required for a stated increase in release volume, then state the condition under which the group must stop raising volume. **Deliverable:** A verification capacity plan with named verifier ownership, per-release evidence requirements, capacity assumptions, and a stop condition. **Independent challenge:** An assurance owner outside the delivery group checks that each proposed check has a trigger, a scope, authority to block, and a recorded place its disposition would land, and rejects any check whose owner also produces the work. **Evidence to retain** - Verifier inventory - Ownership map - Per-release evidence requirements - Capacity assumptions - Stop condition and its measure **Connected knowledge** - [Independent verifier systems](https://aisdlc.ai/agentic-engineering/multi-agent-verification) - [Evidence engineering](https://aisdlc.ai/agentic-engineering/evidence-engineering) - [Eval-driven development](https://aisdlc.ai/agentic-engineering/eval-driven-development) - [Observability & control](https://aisdlc.ai/agentic-engineering/observability-control) - [Human accountability](https://aisdlc.ai/agentic-engineering/human-accountability) **Connected Insights** - [Verification is a plane, not a phase](https://aisdlc.ai/insights/verification-is-a-plane-not-a-phase) - [Evidence before autonomy](https://aisdlc.ai/insights/evidence-before-autonomy) ## 05 · Bind governance to execution **Driving question:** What makes a written policy actually reach the moment an agent is about to act? Connect delegated standing, policy decisions, intervention, and named disposition so written rules reach the point where an action can be refused, held, or escalated. ### Learning objectives - Trace one policy statement from document text to a specific decision point. - Distinguish guidance that advises from a control that can refuse an action. - Bind each agent to a sponsor, owner, scope, and expiry. - Record who intervened, on what basis, and with what disposition. ### Study sequence 1. **[Every agent is a governed principal](/insights/every-agent-is-a-governed-principal)** — read, 14 min. Establish the standing question every policy trace depends on. 2. **[Compare written policy with an enforceable control](/agentic-engineering/runtime-policy-enforcement)** — compare, 20 min. Learn to tell a statement of intent from a decision point that can refuse an action. 3. **[Rehearse a hold, a correction, and a named disposition](/control-plane-lab)** — practice, 35 min. Practise finite intervention steps in a study setting while keeping the simulation boundary explicit. ### Field exercise Take three policy statements your organization already publishes and trace each one to the point where an action could be refused, held, or escalated. **Deliverable:** A policy-to-enforcement trace with agent standing records and rehearsed intervention dispositions. **Independent challenge:** A risk or compliance reviewer outside the delivery group checks each trace for a specific decision point and rejects any statement that only restates intent. **Evidence to retain** - Policy statements as published - Enforcement point per statement - Agent standing records - Intervention rehearsal log - Disposition records **Connected knowledge** - [Agent identity & delegated authority](https://aisdlc.ai/agentic-engineering/agent-identity) - [Runtime policy enforcement](https://aisdlc.ai/agentic-engineering/runtime-policy-enforcement) - [Risk-tiered autonomy](https://aisdlc.ai/agentic-engineering/risk-tiered-autonomy) - [Agent estate governance](https://aisdlc.ai/agentic-engineering/agent-estate-governance) - [Human accountability](https://aisdlc.ai/agentic-engineering/human-accountability) **Connected Insights** - [Every agent is a governed principal](https://aisdlc.ai/insights/every-agent-is-a-governed-principal) - [Evidence before autonomy](https://aisdlc.ai/insights/evidence-before-autonomy) ## 06 · Close the loop from production **Driving question:** How does what production shows change what the organization asks for next? Turn production outcomes, incidents, corrections, and drift into revised intent, durable evaluation cases, and reopened autonomy decisions. ### Learning objectives - Compare the outcome an initiative promised with what production shows. - Convert an incident or correction into a named evaluation case that would have failed before the fix. - Detect drift in behaviour, context sources, and dependencies over time. - Reopen scope and autonomy decisions from operating evidence rather than on a calendar. ### Study sequence 1. **[Signals that describe and signals that decide](/agentic-engineering/observability-control)** — read, 15 min. Separate monitoring that reports the system from signals that should reopen a decision. 2. **[Compare a one-time correction with a standing scope change](/agentic-engineering/continuous-recertification-retirement)** — compare, 20 min. Decide when a finding is a fix and when it should narrow an agent's scope or retire it. 3. **[Convert one production finding into an evaluation case](/assurance-case)** — practice, 35 min. Feed operating evidence back into the standing verification checks. ### Field exercise Take one recorded production incident or missed outcome and carry it back into revised intent, a new evaluation case, and an autonomy decision. **Deliverable:** A learning record linking one production finding to revised intent, a new evaluation case, and an autonomy decision with a named owner. **Independent challenge:** A reviewer independent of the correction runs the new evaluation case against the original behaviour and confirms it fails before the change and passes after. **Evidence to retain** - Production finding - Root-cause note - New evaluation case - Revised intent statement - Autonomy decision record - Open-finding register **Connected knowledge** - [Agent incident response](https://aisdlc.ai/agentic-engineering/agent-incident-response) - [Continuous recertification & retirement](https://aisdlc.ai/agentic-engineering/continuous-recertification-retirement) - [Observability & control](https://aisdlc.ai/agentic-engineering/observability-control) - [Eval-driven development](https://aisdlc.ai/agentic-engineering/eval-driven-development) - [Evidence engineering](https://aisdlc.ai/agentic-engineering/evidence-engineering) - [Intent engineering](https://aisdlc.ai/agentic-engineering/intent-engineering) **Connected Insights** - [Evidence before autonomy](https://aisdlc.ai/insights/evidence-before-autonomy) - [Verification is a plane, not a phase](https://aisdlc.ai/insights/verification-is-a-plane-not-a-phase) ## Capstone · AI-native delivery operating review Produce one operating review for a real delivery group. Measure the whole system, locate the absorption constraint from recorded wait time, redesign batch size and decision packets, size verification and evidence against expected release volume, trace three policy statements to enforcement points, and carry one production finding back into intent. Close the review with a named human disposition and the conditions that would reverse it. ### Deliverables - Balanced measurement set with definitions, collection sources, and known blind spots - Value-stream map with wait time, rework counts, and one named constraint - Revised intake design with batch rules, work-in-process limits, and a decision packet template - Verification and evidence capacity plan with named verifier ownership and a stop condition - Policy-to-enforcement trace with agent standing records and rehearsed intervention dispositions - Production learning record linking one finding to revised intent and a new evaluation case - Named human disposition stating the decision, its basis, and what would reverse it ### Verification - Every figure in the review traces to a named source system and a collection date - The named constraint is supported by recorded wait time and rework counts rather than assertion - Verification ownership is independent of the people and agents generating the work - Every proposed control is labelled as a design proposal, not as an operating control - The final disposition is recorded by a named human decision owner rather than the learner's agent, and states the conditions that would reverse it Progress and completion are self-directed learning records. They are not certification, professional standing, production evidence, or authorization to deploy an agentic system. ## Boundaries This curriculum is original AISDLC editorial synthesis informed by cited research, official documentation, standards, open-source references, and attributed practitioner perspectives. It is learning material—not a standard, production approval, compliance determination, or proof that a depicted control is implemented. Progress and completion are self-directed learning records. They are not certification, professional standing, production evidence, or authorization to deploy an agentic system.